Deep Self-Supervised Learning Network for Multiexposure Image Fusion in Fringe Structured-Light 3-D Reconstruction

Fringe structured-light 3-D reconstruction is a key technology for obtaining the surface morphology of objects in high-precision industrial production. However, in high dynamic range (HDR) scenarios, due to complex object surface reflectivity and lighting conditions, the captured images often contain abnormal exposure areas, leading to poor reconstruction results. To address this challenge, we propose a fringe structured-light multiexposure image fusion (MEF) algorithm based on deep self-supervised learning. We design a self-supervised learning network, SMEF, to fuse images captured under different exposure conditions, achieving fusion without relying on ground truth images and eliminating abnormal exposure regions. Additionally, to mitigate the spectral bias phenomenon where neural networks tend to learn low-frequency information during image fusion while preserving the structured-light fringe characteristics, we designed a combined loss function based on structural similarity and Fourier domain losses to enhance the structured-light fringes. Experimental results demonstrate that the fused images not only improve the accuracy and robustness to noise of 3-D reconstruction but also produce a more complete reconstruction, effectively addressing the issue of point cloud gaps in structured-light 3-D reconstruction in HDR scenarios.